harness-engineering-bootstrap-greenfield

Set up a fresh repo for AI coding agents in ~2–3 hours, in 10 commits. This skill operationalizes section 10 of the source repo — that section is the source of truth for the why; this skill is the how you actually run it.

Loading discipline (progressive disclosure)

Only this SKILL.md is always loaded. Every other file in this skill folder is loaded on demand. Do not pre-fetch. Concretely:

• Step 2 loads only scripts/check-greenfield.sh (for execution, not reading).
• Step 3 loads only assets/ROADMAP.md.template.
• Step 4 loads references/shape-universal.md (always), plus exactly one references/stacks/<STACK>.md matching the detected stack, plus exactly one references/repo-shapes/<SHAPE>.md matching the detected shape. Do not read other stacks or shapes.
• Per commit in Step 4, load only the assets/ template(s) needed for that specific commit. Do not bulk-read the assets/stacks/<STACK>/ directory.
• references/first-feature.md, references/first-year-matrix.md, and references/second-engineer.md are never read by this skill — they are surfaced as links in Step 5 for the user to invoke separately.
• references/quality-rubric.md is loaded only at the gate after Commit 6 and at final validation after Commit 10 (see Step 7). Load only the subsection for the current phase.
• evals/evals.json and scripts/lint-content.sh are loaded only at Step 7.

This keeps the always-loaded surface to ~3KB. Loading every asset eagerly would be 10×+ that and would burn context the agent needs for actual reasoning.

Workflow